A New Memory Chip Survives 700°C and Could Enable AI in Space
Heat has never been a friend to electronic components, and memory chips are no exception. Most chips begin to fail at temperatures above roughly 200 degrees Celsius, limiting their use in extreme temperature scenarios like space exploration and industrial sensing. Researchers at the University of Southern California now say they have created a device that pushes far beyond that boundary, operating at 700 degrees Celsius (1300 degrees Fahrenheit) without failure.
In a paper recently published in the journal Science, a team led by USC engineer Joshua Yang has introduced a new type of memory device that can function at temperatures hotter than molten lava, with the device showing no indication of degradation at that threshold. And it’s possible the device could withstand even hotter temperatures. According to the researchers, 700 degrees was the upper limit of their testing equipment.
(IM Imagery/Shutterstock)
The device is a nanoscale component known as a memristor, which can both store data and perform certain types of computation. That dual role has made memristors a hot topic in AI hardware, where moving data between memory and compute remains a major source of latency and energy consumption. Structurally, the chip consists of a layered stack of tungsten, hafnium oxide, and graphene. Tungsten, which forms the top electrode, has the highest melting point of any metal. Hafnium oxide, a ceramic material already used in semiconductor manufacturing, serves as the insulating layer. At the base is graphene, a single layer of carbon atoms known for its thermal stability and its weak interaction with metal atoms.
That bottom graphene layer is the primary innovation. In conventional devices, high temperatures cause metal atoms from the top electrode to migrate through the insulating layer. Over time, they form a conductive filament that permanently connects the electrodes and short-circuits the device. This metal migration mechanism is a common way these devices fail under high temperatures.
In the USC design, graphene disrupts that process. Tungsten atoms that reach the graphene interface interact very weakly with it and do not readily adhere or cluster. Without a stable anchor point, the formation of a conductive filament is suppressed, allowing the device to maintain its switching behavior even under extreme thermal stress. The researchers confirmed this mechanism through a combination of electron microscopy, spectroscopy, and quantum-level simulations. Their analysis suggests the effect results from how the tungsten atoms interact with the graphene surface, indicating that other materials with similar surface characteristics could potentially offer comparable stability.
Space-based data centers could perform real-time data processing in space (Shutterstock AI)
Even at 700 degrees Celsius, the chip retains data, switches reliably, and operates at low voltage. The device retained data for more than 50 hours without refreshing, endured over one billion switching cycles, and operated at low voltage with switching speeds in the tens of nanoseconds.
Beyond its ability to operate at extreme temperatures, the chip could open up new possibilities for AI computing. Memristors can perform matrix multiplication directly in hardware using Ohm’s Law, where voltage and conductance produce a current that represents the result of the computation. Instead of moving data back and forth between memory and processors, the calculation happens where the data is stored. That approach could reduce both the energy use and latency for linear algebra-heavy workloads.
While this study focuses on high-temperature operation, Yang’s group has already demonstrated memristor-based chips for machine learning at room temperature through its startup, TetraMem. The high-temperature version extends that concept into environments where conventional AI hardware cannot operate, opening the possibility of running inference directly on spacecraft, probes, or industrial systems.
Directly running inference on spacecraft is something Nvidia CEO Jensen Huang touted at GTC when introducing the company’s Vera Rubin Space-1 Module, a compact AI computing system designed to operate aboard satellites and future orbital data centers, where it would process data from space-based sensors in real time rather than sending it back to Earth.
“In space, there’s no conduction, there’s no convection, there’s just radiation. So, we have to figure out how to cool these systems out in space,” Huang said during his keynote.
Nvidia CEO Jensen Huang introduced the Vera Rubin Space-1 Module at GTC
The cooling constraint is just one challenge of running AI systems in space, and components like the USC memristor that can continue operating under extreme thermal conditions could be a solution. There are many other scenarios where this technology could be helpful, like in the energy sector. High-temperature electronics are a requirement in geothermal energy systems, where sensors must operate deep underground, as well as in nuclear and fusion environments. Even in less extreme settings, devices rated for these conditions could improve the reliability of automotive and industrial systems that regularly operate near their thermal limits.
This research is still at an early stage. The current device represents only the memory component of the computing architecture, and additional work will be required to develop other aspects, like high-temperature logic circuits. The current chip prototypes were built by hand at a small scale in a laboratory setting, so design and manufacture will require more time. Still, the materials themselves offer a good starting point. Tungsten and hafnium oxide are already widely used in semiconductor production, and while graphene is less mature, it is advancing toward wafer-scale fabrication.
This work, carried out through a multi-institution collaboration led by USC and supported by the Air Force, changes the conversation around what is technically feasible. For the first time, a key piece of computing hardware has been shown to operate under conditions that were previously out of reach, opening the door to systems that can boldly compute where no systems have computed before.
Related
A New Memory Chip Survives 700°C and Could Enable AI in Space
Heat has never been a friend to electronic components, and memory chips are no exception….
CAS Launches ‘Newton’ Agentic AI Built on Curated Scientific Data
COLUMBUS, Ohio, April 9, 2026 — CAS, a division of the American Chemical Society, has…
Anthropic Unveils ‘Project Glasswing’ as Claude Mythos Targets Software Vulnerabilities
April 9, 2026 — Anthropic has announced Project Glasswing, a new initiative that brings together Amazon…
Linux Foundation A2A Protocol Marks One Year with Broad Enterprise and Cloud Adoption
SAN FRANCISCO, April 9, 2026 — The A2A (Agent-to-Agent) Protocol project, hosted by the Linux…
SiFive Raises $400M Series G to Advance RISC-V Architecture for AI Infrastructure
SANTA CLARA, Calif., April 9, 2026 — SiFive today announced it has raised $400 million…
CIQ Rolls Out ARCA Single-Stack Solution for Academic HPC Operations
RENO, Nev., April 9, 2026 — CIQ, the founding support and services partner of Rocky…
